Preliminary approach to predict the reactivation of long-term kinematics landslides through noval synergistic stacked deep learning approach

Landslides pose serious risks to both natural landscapes and urban infrastructure, often triggered by complex interactions between geological conditions and meteorological events such as intense rainfall. This study presents a novel stacked deep learning framework that integrates Graph Convolutional Networks (GCN) with GCN-based Long Short-Term Memory (GCN-LSTM) models to improve the prediction of landslide-induced surface deformation. The case study focuses on the Randazzo Landslide in northeastern Sicily, a region with intricate geological structures and recurrent landslide events. We utilize high-resolution satellite radar data from the COSMO-SkyMed mission, along with comprehensive geological, geomorphological, and rainfall datasets, to capture the spatial and temporal patterns governing landslide behavior. The spatial component of the model leverages GCN to extract non-Euclidean spatial relationships among predisposing factors, while the temporal component applies GCN-LSTM to model the progression of rainfall and ground deformation over time, as obtained through Multi-temporal Interferometric Synthetic Aperture Radar analysis. Outputs from both base models are fed into a GCN-based meta-model, which synthesizes these features to enhance prediction accuracy. The framework was trained and validated on data collected between 2011 and 2014, demonstrating strong predictive performance in terms of Mean Absolute Error, Root Mean Squared Error, and R-squared metrics. Results indicate that the stacked model outperforms standalone GCN and GCN-LSTM implementations. This methodology provides a scalable, adaptable tool for forecasting landslide deformation and contributes to the advancement of early warning systems and hazard management strategies through the fusion of remote sensing data and advanced deep learning techniques.